Questions: Bias-Complexity Tradeoff (Formal)

4 questions to test your understanding

Score: 0 / 4
Question 1 Multiple Choice

You train a linear classifier (VC dimension 3) and a degree-10 polynomial classifier (VC dimension 11) on 50 data points. The polynomial achieves lower training error but higher test error. How does the bias-complexity decomposition explain this?

AThe polynomial has lower approximation error (it can fit the target better) but much higher estimation error (VC dimension 11 with only 50 points means the ERM solution is unreliable), and the estimation error dominates
BThe polynomial has higher approximation error because degree-10 polynomials introduce systematic distortion
CBoth have the same approximation error, but the polynomial's training algorithm is less efficient
DThe polynomial has lower estimation error because it fits the training data better, but higher bias from the polynomial assumption
Question 2 True / False

In the bias-complexity decomposition, increasing the hypothesis class size ALWAYS reduces approximation error.

TTrue
FFalse
Question 3 True / False

A model with zero approximation error and very high estimation error will generally perform worse than a model with moderate approximation error and low estimation error.

TTrue
FFalse
Question 4 Short Answer

How does the bias-complexity tradeoff differ from the classical bias-variance tradeoff, and what does the formal version add?

Think about your answer, then reveal below.